<p><strong>TLDR:</strong> Meta is laying off roughly 8,000 people on May 20 as it reorganizes around AI pods, funded by a $115–135B capex plan — not financial distress. Claude Opus 4.7 quietly shipped with a 1.46x token inflation for text and up to 3x for images at the same per-token price as 4.6, which is a meaningful cost bump that most users haven’t noticed yet. Google’s Aletheia solved 6 of 10 novel research-level math problems zero-shot, edging out OpenAI’s supervised attempt on the same benchmark. Vercel got breached and API keys are reportedly being sold for $2M, which is a bad Sunday for anyone who vibe-coded their way to production.</p>
<hr>
<h2>Meta lays off 8,000 to fund AI infrastructure</h2>
<p>Meta will cut approximately 8,000 employees — about 10% of its workforce — on May 20, with more cuts planned for the second half of 2026. This isn’t a distress move: the company posted $201B in revenue and $43.6B in free cash flow in 2025. The restructuring is about redirecting headcount budget toward a $115–135B AI infrastructure spend, with teams reorganized into AI-focused pods under new Chief AI Officer Alexandr Wang. Traditional roles are being replaced with titles like “AI builder” and “AI pod lead.” Since 2022, Zuckerberg has cut roughly 25,000 positions total. The pattern is consistent across the industry: record revenues, shrinking headcount, massive infrastructure bets.</p>
<h2>Opus 4.7 uses significantly more tokens than 4.6</h2>
<p>One developer upgraded their Claude token counter tool to compare models and found Opus 4.7 uses 1.46x the tokens for text and up to 3x for images compared to 4.6 — at the same per-token price. That’s a real price increase disguised as a model upgrade. Early users report genuine quality improvements, particularly on UI tasks and workflow reliability (Descript saw 18% lower cost and 25% faster performance in their evals, which suggests their workload skews text-heavy). If your usage is image-heavy, run your own numbers before assuming the upgrade is cost-neutral.</p>
<h2>Google’s Aletheia solves novel math problems autonomously</h2>
<p>Google DeepMind’s Aletheia, built on Gemini 3 Deep Think, solved 6 of 10 unpublished research-level math problems in the FirstProof challenge with zero human hints. It uses a multi-agent loop of Generator, Verifier, and Reviser components plus search, and outputs “No solution found” rather than hallucinating — which the researchers treat as a feature, not a limitation. OpenAI attempted the same benchmark with an unreleased model using limited human supervision, initially claimed 6 solutions, then revised to 5 after one was found flawed. Aletheia also scored ~91.9% on IMO-ProofBench. The researchers are honest that it still makes more errors than human experts, but the zero-shot, no-hallucination result is a meaningful data point.</p>
<h2>Vercel breach puts API keys on the market</h2>
<p>A security incident at Vercel resulted in stolen API keys reportedly being sold for $2M. The breach involved a third-party AI tool (identified as Context) and a chain of vulnerabilities including prompt injection via GitHub issue titles into an AI-powered triage bot with Bash/Write tool access, followed by GitHub Actions cache poisoning to exfiltrate secrets. If you have any Vercel-connected projects, rotate everything now. The broader pattern here is worth paying attention to: AI coding tools that interact with environment variables and CI runners are expanding the attack surface in ways most teams haven’t fully mapped.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Grok 4.3 beta</strong> is live for SuperGrok and X Premium+ users, with Skills support already embedded in system instructions — hidden from users for now but apparently functional.</li>
<li><strong>GPT-5.4 and Codex</strong> are getting attention: one tweet from the Codex team hints at “at least an order of magnitude in speedups” coming, and GPT-5.4’s GPU kernel knowledge is drawing genuine praise from ML engineers.</li>
<li><strong>AWS Agent Registry</strong> launched in preview, giving enterprises a centralized way to track and govern AI agents across their infrastructure — a direct response to agent sprawl as orgs scale.</li>
<li><strong>MCP hits 110M monthly downloads</strong>, reaching that milestone in roughly half the time React did, with a stateless transport protocol from Google landing in June and TypeScript/Python SDK v2 on the roadmap.</li>
<li><strong>Vibe coding security debt</strong> is compounding: one developer reports routinely finding serious data leaks in AI-generated apps without looking for them — entire user databases returned to every client, plaintext passwords on unauthenticated endpoints, authorization logic on the frontend only.</li>
<li><strong>Claude Design</strong> launched quietly in research preview for Pro/Max/Team/Enterprise subscribers, with a conversation-to-canvas loop and direct Claude Code integration; Figma’s stock dropped ~7% on launch day.</li>
<li><strong>Codex desktop computer use</strong> added an Expo plugin, putting it in a more practical position for React Native developers beyond one-off code generation.</li>
<li><strong>Anthropic confirmed</strong> an erroneous account suspension affecting T3 Code users; the account has been reinstated.</li>
<li><strong>GitHub Copilot can now run locally via Ollama</strong> with no API costs, which removes one of the main friction points for teams that want local inference.</li>
<li><strong>Developer ramp-up time</strong> dropped from 91 days to 33 days (time to 10th PR) at AI-heavy teams, per data cited in an engineering management newsletter — though those same teams are merging twice as many PRs and spending 91% more time reviewing them.</li>
</ul>


AI coding assistants, code generation tools, and vibe coding workflows.

Agentic Digest

A developer-focused news roundup covering: Meta's 8,000-person layoff to fund $115–135B AI infrastructure (not financial distress); Claude Opus 4.7's hidden token cost increase (1.46x text, up to 3x images vs 4.6 at same per-token price); Google DeepMind's Aletheia solving 6/10 novel research math problems zero-shot; and a Vercel security breach involving prompt injection and GitHub Actions cache poisoning that put API keys on sale for $2M. Also covers Grok 4.3 beta, MCP hitting 110M monthly downloads, GitHub Copilot running locally via Ollama, Claude Design launch, and growing security debt in AI-generated code.

Meta cuts 8,000 jobs to fund AI infrastructure, Opus 4.7 ships with a hidden token cost

<p>Please please please add links to sources.</p>


<p>Vercel incident:
<a href="https://vercel.com/kb/bulletin/vercel-april-2026-security-incident" target="_blank" rel="noopener nofollow">https://vercel.com/kb/bulletin/vercel-april-2026-security-incident</a></p>
